Goroutine Stack Buffer

Saya bertanya-tanya dengan menempatkan array dan irisan pada stack, seperti std :: array di C ++ dan array di C. Tumpukan itu ditulis dengan baik oleh Vincent Blanchon dalam artikel Go: Bagaimana Ukuran Tumpukan Goroutine Berkembang? . Vincent berbicara tentang menumpuk perubahan di goroutine. Singkatnya:

• ukuran tumpukan minimum 2 KB;
• ukuran maksimum tergantung pada arsitektur, pada arsitektur 64-bit 1 GB;
• setiap kali ukuran tumpukan berlipat ganda;
• tumpukan meningkat dan menurun.

Saya akan mencari tahu berapa banyak yang bisa diletakkan di tumpukan dan biaya overhead apa yang bisa dihasilkan.Saya akan menggunakan go 1.13.8 yang dikompilasi dari sumber dengan informasi debug dan opsi kompilasi -gcflags=-m.

Untuk mendapatkan hasil debug saat program sedang berjalan, Anda harus runtime/stack.gomenetapkan konstanta stackDebugke nilai yang diinginkan:

const (
    // stackDebug == 0: no logging
    //            == 1: logging of per-stack operations
    //            == 2: logging of per-frame operations
    //            == 3: logging of per-word updates
    //            == 4: logging of per-word reads
    stackDebug       = 0
    //...
)

Program yang sangat sederhana

Saya akan mulai dengan sebuah program yang tidak melakukan apa-apa untuk melihat bagaimana tumpukan menonjol di bawah goroutine utama:

package main

func main() {
}

Kami melihat bagaimana kumpulan global diisi dengan segmen 2, 8, 32 KB:

$ GOMAXPROCS=1 ./app 
stackalloc 32768
  allocated 0xc000002000
stackalloc 2048
stackcacherefill order=0
  allocated 0xc000036000
stackalloc 32768
  allocated 0xc00003a000
stackalloc 8192
stackcacherefill order=2
  allocated 0xc000046000
stackalloc 2048
  allocated 0xc000036800
stackalloc 2048
  allocated 0xc000037000
stackalloc 2048
  allocated 0xc000037800
stackalloc 32768
  allocated 0xc00004e000
stackalloc 8192
  allocated 0xc000048000

Ini poin penting. Kita dapat berharap bahwa memori yang dialokasikan akan digunakan kembali.

bukan program yang sangat sederhana

Di mana tanpa Hello world !

package main

import "fmt"

func helloWorld()  {
    fmt.Println("Hello world!")
}

func main() {
    helloWorld()
}

runtime: newstack sp=0xc000036348 stack=[0xc000036000, 0xc000036800]
    morebuf={pc:0x41463b sp:0xc000036358 lr:0x0}
    sched={pc:0x4225b1 sp:0xc000036350 lr:0x0 ctxt:0x0}
stackalloc 4096
stackcacherefill order=1
  allocated 0xc000060000
copystack gp=0xc000000180 [0xc000036000 0xc000036350 0xc000036800] -> [0xc000060000 0xc000060b50 0xc000061000]/4096
stackfree 0xc000036000 2048
stack grow done
stackalloc 2048
  allocated 0xc000036000
Hello world!

Ukuran 2 KB tidak cukup dan alokasi 4 KB blok pergi. Kami akan berurusan dengan angka dalam tanda kurung. Ini adalah alamat, di tepinya, ini adalah nilai struktur:

// Stack describes a Go execution stack.
// The bounds of the stack are exactly [lo, hi),
// with no implicit data structures on either side.
type stack struct {
    lo uintptr
    hi uintptr
}

Di tengah, ini adalah alamat yang menunjukkan berapa banyak tumpukan digunakan, dihitung sebagai berikut:

//...
used := old.hi - gp.sched.sp
//...
print("copystack gp=", gp, " [", hex(old.lo), " ", hex(old.hi-used), " ", hex(old.hi), "]", " -> [", hex(new.lo), " ", hex(new.hi-used), " ", hex(new.hi), "]/", newsize, "\n")

:

- :

package main

import (
    "fmt"
    "time"
)

func helloWorld()  {
    fmt.Println("Hello world!")
}

func main() {
    for i := 0 ; i< 12 ; i++ {
        go helloWorld()
    }

    time.Sleep(5*time.Second)
}

$ ./app 2>&1 | grep "alloc 4"
stackalloc 4096
stackalloc 4096
stackalloc 4096
stackalloc 4096
stackalloc 4096

, . .

stack_test.go , . bigsize, go build -gcflags=-m main.go:8:6: moved to heap: x.

package main

const bigsize = 1024*1024*10
const depth = 50

//go:noinline
func step(i int) byte {
    var x [bigsize]byte
    if i != depth {
        return x[i] * step(i+1)
    } else {
        return x[i]
    }
}

func main() {
    step(0)
}

10 MB, depth > 50 stack overflow:

runtime: goroutine stack exceeds 1000000000-byte limit
fatal error: stack overflow

, -gcflags="-m"

package main

import (
    "fmt"
    "os"
)

const bigsize = 1024 * 1024 * 10

func readSelf(buff []byte) int { // readSelf buff does not escape
    f, err := os.Open("main.go") // inlining call to os.Open
    if err != nil {
        panic(err)
    }

    count, _ := f.Read(buff)
    if err := f.Close(); err != nil { // inlining call to os.(*File).Close
        panic(err)
    }

    return count
}

func printSelf(buff []byte, count int) { // printSelf buff does not escape
    tmp := string(buff[:count]) // string(buff[:count]) escapes to heap
    // inlining call to fmt.Println
    // tmp escapes to heap
    // printSelf []interface {} literal does not escape
    // io.Writer(os.Stdout) escapes to heap
    fmt.Println(tmp)
}

func foo() {
    var data [bigsize]byte
    cnt := readSelf(data[:])
    printSelf(data[:], cnt)
}

func main() {
    foo() // can inline main
}

, , :

stackalloc 2048
stackcacherefill order=0
  allocated 0xc000042000
runtime: newstack sp=0xc00008ef40 stack=[0xc00008e000, 0xc00008f000]
    morebuf={pc:0x48e4e0 sp:0xc00008ef50 lr:0x0}
    sched={pc:0x48e4b8 sp:0xc00008ef48 lr:0x0 ctxt:0x0}
stackalloc 8192
stackcacherefill order=2
  allocated 0xc000078000
copystack gp=0xc000000180 [0xc00008e000 0xc00008ef48 0xc00008f000] -> [0xc000078000 0xc000079f48 0xc00007a000]/8192
stackfree 0xc00008e000 4096
stack grow done
...
runtime: newstack sp=0xc0010aff40 stack=[0xc0008b0000, 0xc0010b0000]
    morebuf={pc:0x48e4e0 sp:0xc0010aff50 lr:0x0}
    sched={pc:0x48e4b8 sp:0xc0010aff48 lr:0x0 ctxt:0x0}
stackalloc 16777216
  allocated 0xc0010b0000
copystack gp=0xc000000180 [0xc0008b0000 0xc0010aff48 0xc0010b0000] -> [0xc0010b0000 0xc0020aff48 0xc0020b0000]/16777216
stackfree 0xc0008b0000 8388608
stack grow done

, .

64 KB :

package main

import (
    "fmt"
    "os"
)

const bigsize = 1024*64 - 1

func readSelf(buff []byte) int { // readSelf buff does not escape
    f, err := os.Open("main.go") // inlining call to os.Open
    if err != nil {
        panic(err)
    }

    count, _ := f.Read(buff)
    if err := f.Close(); err != nil { // inlining call to os.(*File).Close
        panic(err)
    }

    return count
}

func printSelf(buff []byte, count int) { // printSelf buff does not escape
    tmp := string(buff[:count]) // string(buff[:count]) escapes to heap
    // inlining call to fmt.Println
    // tmp escapes to heap
    // printSelf []interface {} literal does not escape
    // io.Writer(os.Stdout) escapes to heap
    fmt.Println(tmp)
}

func foo() {
    data := make([]byte, bigsize) // make([]byte, bigsize) does not escape
    cnt := readSelf(data)
    printSelf(data, cnt)
}

func main() { // can inline main
    foo()
}

stackalloc 131072
  allocated 0xc0000d0000
copystack gp=0xc000000180 [0xc0000c0000 0xc0000cff48 0xc0000d0000] -> [0xc0000d0000 0xc0000eff48 0xc0000f0000]/131072
stackfree 0xc0000c0000 65536

Go (zero values). . , sync.Pool.

:

• ( );
• ;
• ( /dev/null).

, , .

func readSelf(buff []byte) int {
    f, err := os.Open("bench_test.go")
    if err != nil {
        panic(err)
    }

    count, err := f.Read(buff)
    if err != nil {
        panic(err)
    }

    if err := f.Close(); err != nil {
        panic(err)
    }

    return count
}

func printSelf(buff []byte, count int) {
    f, err := os.OpenFile(os.DevNull, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
    if err != nil {
        panic(err)
    }
    if _, err := f.Write(buff[:count]); err != nil {
        panic(err)
    }

    if err := f.Close(); err != nil {
        panic(err)
    }
}

//go:noinline
func usePoolBlock4k() {
    inputp := pool4blocks.Get().(*[]byte)
    outputp := pool4blocks.Get().(*[]byte)

    cnt := readSelf(*inputp)
    copy(*outputp, *inputp)
    printSelf(*outputp, cnt)

    pool4blocks.Put(inputp)
    pool4blocks.Put(outputp)
}

//go:noinline
func useStackBlock4k() {
    var input [block4k]byte
    var output [block4k]byte

    cnt := readSelf(input[:])
    copy(output[:], input[:])
    printSelf(output[:], cnt)
}

func BenchmarkPoolBlock4k(b *testing.B) {
    runtime.GC()
    for i := 0; i < 8; i++ {
        data := pool4blocks.Get()
        pool4blocks.Put(data)
    }
    for i := 0; i < 8; i++ {
        usePoolBlock4k()
    }

    b.ResetTimer()

    for i := 0; i < b.N; i++ {
        usePoolBlock4k()
    }
}

func BenchmarkStackBlock4k(b *testing.B) {
    runtime.GC()
    for i := 0; i < 8; i++ {
        useStackBlock4k()
    }
    b.ResetTimer()

    for i := 0; i < b.N; i++ {
        useStackBlock4k()
    }
}

$ go test -bench=. -benchmem
goos: linux
goarch: amd64
pkg: gostack/04_benchmarks
BenchmarkPoolBlock4k-12           312918          3804 ns/op         256 B/op          8 allocs/op
BenchmarkStackBlock4k-12          313255          3833 ns/op         256 B/op          8 allocs/op
BenchmarkPoolBlock8k-12           300796          3980 ns/op         256 B/op          8 allocs/op
BenchmarkStackBlock8k-12          294266          4110 ns/op         256 B/op          8 allocs/op
BenchmarkPoolBlock16k-12          288734          4138 ns/op         256 B/op          8 allocs/op
BenchmarkStackBlock16k-12         269382          4408 ns/op         256 B/op          8 allocs/op
BenchmarkPoolBlock32k-12          272139          4407 ns/op         256 B/op          8 allocs/op
BenchmarkStackBlock32k-12         240339          4957 ns/op         256 B/op          8 allocs/op
PASS

4 KB, . , . :

(pprof) top 15
Showing nodes accounting for 7.99s, 79.58% of 10.04s total
Dropped 104 nodes (cum <= 0.05s)
Showing top 15 nodes out of 85
      flat  flat%   sum%        cum   cum%
     3.11s 30.98% 30.98%      3.38s 33.67%  syscall.Syscall6
     2.01s 20.02% 51.00%      2.31s 23.01%  syscall.Syscall
     0.83s  8.27% 59.26%      0.83s  8.27%  runtime.epollctl
     0.60s  5.98% 65.24%      0.60s  5.98%  runtime.memmove
     0.21s  2.09% 67.33%      0.21s  2.09%  runtime.unlock
     0.18s  1.79% 69.12%      0.18s  1.79%  runtime.nextFreeFast
     0.14s  1.39% 70.52%      0.33s  3.29%  runtime.exitsyscall
     0.14s  1.39% 71.91%      0.21s  2.09%  runtime.reentersyscall
     0.13s  1.29% 73.21%      0.46s  4.58%  runtime.mallocgc
     0.12s  1.20% 74.40%      1.25s 12.45%  gostack/04_benchmarks.useStackBlock32k
     0.12s  1.20% 75.60%      0.13s  1.29%  runtime.exitsyscallfast
     0.11s  1.10% 76.69%      0.45s  4.48%  runtime.SetFinalizer
     0.11s  1.10% 77.79%      0.11s  1.10%  runtime.casgstatus
     0.10s     1% 78.78%      0.13s  1.29%  runtime.deferreturn
     0.08s   0.8% 79.58%      1.24s 12.35%  gostack/04_benchmarks.useStackBlock16k

syscalls. useStackBlock32k getFromStack32k. 120ms:

(pprof) list useStackBlock32k
Total: 10.04s
ROUTINE ======================== useStackBlock32k in bench_test.go
     120ms      1.25s (flat, cum) 12.45% of Total
         .          .    158:   printSelf(output[:], cnt)
         .          .    159:}
         .          .    160:
         .          .    161://go:noinline
         .          .    162:func useStackBlock32k() {
      90ms       90ms    163:   var input [block32k]byte
      30ms       30ms    164:   var output [block32k]byte
         .          .    165:
         .      530ms    166:   cnt := readSelf(input[:])
         .      160ms    167:   copy(output[:], input[:])
         .      440ms    168:   printSelf(output[:], cnt)
         .          .    169:}

ROUTINE ======================== usePoolBlock32k in bench_test.go
      10ms      1.18s (flat, cum) 11.75% of Total
         .          .    115:   pool16blocks.Put(outputp)
         .          .    116:}
         .          .    117:
         .          .    118://go:noinline
         .          .    119:func usePoolBlock32k() {
         .       10ms    120:   inputp := pool32blocks.Get().(*[]byte)
         .       10ms    121:   outputp := pool32blocks.Get().(*[]byte)
         .          .    122:
         .      520ms    123:   cnt := readSelf(*inputp)
      10ms      200ms    124:   copy(*outputp, *inputp)
         .      440ms    125:   printSelf(*outputp, cnt)
         .          .    126:
         .          .    127:   pool32blocks.Put(inputp)
         .          .    128:   pool32blocks.Put(outputp)
         .          .    129:}

:

• 10 MB;
• 64 KB;
• ada overhead untuk memecah tumpukan;
• Ada overhead di nilai nol .